MacBook M5 Pro and Qwen3.5 = Local AI Security System

Qwen3.5-9B running locally on a MacBook Pro M5 scores 93.8% on HomeSec-Bench, within 4 points of GPT-5.4, at 25 tok/s using only 13.8 GB unified memory with zero API costs. HomeSec-Bench is a 96-test evaluation across 16 suites designed to assess LLMs on real home security workflows including security classification, event deduplication, and tool use. Local models also demonstrate competitive latency, with Qwen3.5-35B-MoE achieving lower TTFT than all OpenAI cloud models.